Inductive Dual-Polarity Modeling via Static-Dynamic Disentanglement for Dynamic Signed Networks

A research team led by Yikang Hou and Junjie Huang has introduced IDP-DSN (Inductive Dual-Polarity framework for Dynamic Signed Networks), a novel AI model designed to predict evolving relationships in networks where connections can be positive (trust) or negative (conflict). Unlike previous methods that mix positive and negative signals in shared representations, IDP-DSN maintains separate "memories" for each polarity and disentangles static user traits from dynamic behavioral patterns. This approach specifically addresses the cold-start problem where the model must make predictions about users it has never seen during training.

The framework was tested on real-world datasets including Bitcoin trading networks (BitcoinAlpha, BitcoinOTC), Wikipedia administrator elections (Wiki-RfA), and product review platforms (Epinions). In inductive settings—where test edges involve at least one completely new user—IDP-DSN achieved remarkable performance gains of 16.9-30.1% in Macro-F1 scores over previous state-of-the-art methods. For instance, on the BitcoinOTC dataset, it showed 24% improvement in inductive evaluation, demonstrating superior ability to generalize from limited historical data about new participants.

Key innovations include sign-selective temporal modeling that tracks positive and negative interaction patterns independently, and a history-only neighborhood inference mechanism that builds representations for unseen nodes solely from their past interactions rather than requiring pre-learned embeddings. The model's architecture allows it to forecast whether future relationships between users will be cooperative, adversarial, or non-existent, making it particularly valuable for platforms dealing with trust dynamics, moderation, or fraud detection.